Deepgram Flux

Audio modelCard summarized

The first real-time Conversational Speech Recognition model built for voice agents.

● Re-checked 12 Oct 2026 · 62 facts · 0 corrections · 3 unconfirmedSource: Flux launch post (Deepgram publishes no model card)

Use it via API

$0.0077 / min

Code example Compare

Where to call it

Deepgram APIflux-general-en

Supports

Streaming

Full feature support

Good for

Real-time voice agents and conversational apps
Latency-sensitive agents that start LLM calls speculatively on EagerEndOfTurn
Voice agents, IVR, agent assist and real-time transcription

Overview

Deepgram · 02Flux

Flux

The first real-time Conversational Speech Recognition model built for voice agents.

Typical price

$0.0077 / min

Flux English, streaming

Max length: —
Languages: English (flux-general-en); 10 languages in flux-general-multi: English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, Dutch
Modes: speech-to-text, streaming
Released: 1 Oct 2025

Flux is Deepgram's conversational speech recognition model for voice agents: the same model that transcribes also detects end of turn, and it streams turn-based transcripts over /v2/listen. Deepgram claims EndOfTurn fires within 1.5 s 95% of the time, about 30% fewer false interruptions than pipeline approaches, and WER nearly identical to Nova-3 streaming. It is streaming only; Deepgram points to Nova-3 for pre-recorded audio, meetings and analytics. launch post; Flux vs Nova-3

flux-general-en

Typical price: $0.0077 / min
Languages: English (flux-general-en); 10 languages in flux-general-multi: English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, Dutch
Weights: Closed
Released: 1 Oct 2025

speech-to-text
streaming
end-of-turn detection
eager end-of-turn (speculative LLM start)
keyterm prompting
language hints and detection (multilingual)
number redaction